Graphiti MCP
Graphiti MCP Server
Zep's Graphiti MCP server: long-term agent memory as a time-aware knowledge graph on FalkorDB or Neo4j
Install
git clone https://github.com/getzep/graphiti.git
cd graphiti/mcp_server
docker compose upStarts FalkorDB and the MCP server in one container. Set your LLM provider key first.
This is third-party code. Review the repository files before installing.
What it does
Graphiti builds a knowledge graph that grows from conversations, documents and JSON and keeps the history of how facts change. The MCP server in mcp_server gives the agent tools to add episodes, find facts and entities with semantic and hybrid search, group data by group_id and maintain the graph. Typed entities are extracted from text: preferences, requirements, procedures, organizations, events and more. Episodes are processed in a queue with configurable concurrency.
Who it is for. For developers who need their agent to remember context across sessions and projects.
Good fit when
- The agent needs long-term memory of user preferences and project decisions
- You need to track how facts changed over time
- You already use Neo4j or can run a FalkorDB container
Not a fit when
- A simple notes file or CLAUDE.md is enough
- You cannot afford many LLM calls during ingestion
Example request
Remember that this project uses pnpm and we never touch the legacy folderLimitations
The server is marked experimental. It needs Docker or Python 3.10+ with uv and an LLM provider key, OpenAI by default; Ollama is supported. Each episode takes several model calls, and small local models often break structured output.
How to disable. Stop the docker compose containers and remove the graphiti-memory server from your MCP client config.
MCP
- Transport
- http
- Authentication
- API key
| Environment variables | |
|---|---|
| OPENAI_API_KEY required, secret | Default LLM provider key; another provider needs its own key |
| NEO4J_URI | Neo4j address when used instead of FalkorDB |
| NEO4J_PASSWORD secret | Neo4j password |
| SEMAPHORE_LIMIT | How many episodes to process concurrently |
Security check
- Sends episode content to an external LLM provider
- Runs database containers
README in short
The server README explains how Graphiti differs from plain RAG and lists the MCP features. The quick start uses docker compose with FalkorDB or a separate Neo4j setup, plus a uv-based config for stdio-only clients. It details config.yaml: database, model and embedding providers, entity types and concurrency limit. Apache-2.0 licensed.
FAQ
Which database should I pick?
FalkorDB ships in the same container as the server by default. The README recommends Neo4j for production.
Can I avoid OpenAI?
Yes, Anthropic, Gemini, Groq, Azure OpenAI and OpenAI-compatible endpoints like Ollama are supported.
Related
Open-source personal AI assistant on your own machine: answers in Telegram, Slack, Discord and WhatsApp, extended with skills and plugins
A self-improving agent from Nous Research with a TUI, messaging gateway, cron jobs and skills it writes itself
An open source coding agent for the terminal and desktop with build and plan modes
Open prompt library with a Claude Code plugin, MCP server and CLI: search, fetch and improve prompts and skills from an agent